Claude AI Guide: How Beginners Actually Start [2026]

Most tutorials treat Claude like “the polite ChatGPT.” The quieter fact is that its backbone is Constitutional AI (Bai et al., Dec 2022, arXiv:2212.08073): the model critiques and revises its own answers against an explicit written constitution, using AI feedback far more than human harm labels. That design still shapes refusals and careful rewrites you’ll hit in 2026.

Claude is Anthropic’s assistant family on claude.ai (web, desktop, iOS, Android). This guide is not a “type hello” tour. Constitution in practice. Plan math as of late 2026. A limit-aware first workflow. Traps that burn the shared usage pool. A blunt stay-or-switch split.

What Constitutional AI Actually Changes for You

Anthropic trained Claude to be helpful, honest, and harmless with written principles – not only contractor thumbs-up/thumbs-down. The public method pairs supervised self-critique with RLAIF against that constitution and minimal human harm labels.

Day to day? Clearer objections instead of foggy “I can’t help.” Tight instruction-following on long briefs. Fewer drive-by topic drifts across revisions. The flip side hurts: security, bio, and dual-use coding prompts get blocked harder than on looser models – even when you’re testing kit you own. Rephrase toward authorized work on your systems, or try another model in the picker. Twenty identical retries just torch quota.

Think of the constitution less like a PR slogan and more like an employee handbook the model actually rereads before it answers. You feel it when a risky ask comes back with a structured no – and when a messy brief still comes back on-rails.

Picker reality (names shift; look at what’s live): Haiku for fast cheap volume. Sonnet as the Free default. Opus for harder reasoning and agent-style work on paid. Fable for long-horizon jobs when your plan allows it. On current paid flagships, chat context often lists up to 1M tokens – varies by model and surface (chat vs Code vs Cowork), and older configs still sit lower.

Claude Plans and the Limit Math Beginners Miss

From claude.com/pricing (late 2026 snapshot – verify before you pay):

No fixed message counter. One shared pool across web, mobile, desktop, and Claude Code. Rolling five-hour session window on every plan; paid adds weekly caps. Hit either wall and the UI shows a reset time – or you can spend usage credits on paid and keep going at API-like rates.

Fable access, per the same pricing models table: no on Free; usage-credit gated on Pro; capped at 50% of weekly limits on Max. Heavy context and Code burn that bar faster than short Q&A.

Before a long session, open Settings → Usage. Already at 60%+ weekly and about to paste a repo or run Code? Park cleanup on Haiku/Sonnet. Or open a fresh Project. One greedy agent loop can erase the rest of the week.

Rationing sounds annoying until you notice what it forces: one meaty brief beats twelve vague follow-ups. Weird question – would you write cleaner prompts if every token came out of a visible weekly jar?

Your First Real Claude Workflow (Limit-Aware)

Skip “hi Claude.” This loop teaches Projects, Artifacts, and revision without torching the session.

1. Open claude.ai. Google or email sign-in. Free is enough.
2. New Project in the sidebar. One recurring job name – “Weekly status pack,” “Client X research.” Free tops out at five total, so don’t waste slots on toys. Paste short standing rules: audience, tone, must-include sections, never-do list. Bloated project files reload every turn. Keep them tight.
3. Chat inside the Project. Attach one real file (notes PDF, CSV, last week’s draft). Brief it like a sharp colleague: goal, constraints, output format, definition of done.

Turn the attached meeting notes into a one-page status for leadership.
Audience: non-technical execs. Sections: Decisions, Risks (with owners), Next 7 days.
Max 400 words. Flag anything unclear instead of inventing numbers.
Return as an Artifact I can export.

Artifact lands – edit inline or send surgical fixes (“Risk #2 owner is Maya, not Jordan; cut the jargon in paragraph 3”). Same thread wins. Context stays warm. You don’t re-pay the quota tax explaining the job again.

Reply suddenly thinner, or older turns collapse into a summary? That’s auto context management on paid plans with code execution near the ceiling – it costs extra usage. Save the Artifact. New chat in the same Project. Re-attach only what still matters. Very large first pastes can still throw hard length errors even when the headline window is huge; split upstream.

Loop: Project for standing memory → one dense brief → Artifact deliverable → tight revisions. Next: chain prompts inside one Project; add connectors once you’re on Pro.

Common Pitfalls That Waste Your First Week

Community reports on long Code sessions: context bloats, quality drifts after repeated compaction. Run whatever context diagnostic your surface offers. Trim permanent instruction files. Don’t assume the 1M headline means “paste everything forever.”

People treat Free like a firehose. Complex Artifacts, web search, multi-file analysis chew the five-hour window faster than short Q&A. Approaching-limit warning? Finish the deliverable and stop. Three “quick” new threads is how the weekly bar dies on paid.

Claude vs ChatGPT (and When to Use Each)

Individual tiers both hover near $20/mo. Claude’s usual wins: long-document fidelity, careful writing, agentic coding via Claude Code. ChatGPT’s usual wins: native image/video, mature voice, wider plugin surface. Claude still doesn’t mint photos natively – it drafts Design/Slides/Docs and codes visualizations. Image-first job? Pair tools or you’ll hate the workflow.

Both advertise large context now. Real free caps and product shape still disagree – check each vendor’s pricing page the week you decide.

Live in Claude if your week is briefs, code, and multi-page sources. Keep ChatGPT when you need pictures or one shallow do-everything app. Lots of people run both and route by task.

FAQ

Is Claude free to use?

Yes. Free includes chat, search, files, Artifacts, and Sonnet/Haiku. Limits are real – heavy days hit the five-hour window.

How big is Claude’s context window right now?

Depends on model and surface. Anthropic’s help docs (paid chat, updated through 2026) list up to 1M tokens for current Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5; many older or alternate setups sit at 200K-500K. See the context window article. Part of the window is reserved for the reply. Near the ceiling, auto-summarization can kick in and spend extra usage. 1M is model-specific. Not universal.

Should I start on Free or jump to Pro?

Start Free for one real work week: a single Project, real files, real Artifacts. That surfaces whether session walls interrupt you mid-task.

Move to Pro when you repeatedly hit those walls, need Opus/Code/Cowork/Research, or outgrow five Projects. Max is for people who already know Pro’s ceiling and live in the tool all day. Annual Pro ($200 up front) only makes sense once you’re sure – the monthly rate drops to $17, but you’re committing.

Open claude.ai. One Project for a task you already owe this week. Run the Artifact brief above before another overview tab. You’ll know fast if it fits.